In the projected case, (1.2, 0.5) divided by 2.0 results in (0.6, 0.25) as
effective coordinates. The 0.25 y-coordinate is well within the second row of
texels; that's fine. On the other hand, the 0.6 x-coordinate falls right
between the third and fourth column of texels. The test expects the fourth
column to be selected, but that's very fragile: neither 1.2 nor 0.6 can be
exactly represented as a 32-bit floating-point value, and it only takes a
single ulp to push things to the other side of the edge.
